perf: drop swiglu warp-rows variant for an M=8 block-size rule

- delete the warp-per-row kernel and the (6912,1536) M=2/4/8 dispatch table; under rotated cold weights the warp path is 2-6% slower than CTA reuse at M=2/4, and the table had been tuned against L2-resident timing
- a single CTA-reuse kernel now serves all M in [1, 8]; block size is 256 threads for M in [1, 7] and 128 for M=8, where the shorter shared-memory reduction tree wins
- document in docs/developer/swiglu_benchmark.md that the earlier operator numbers were L2-resident: the fused kernel sits at the dual-stream cold-read floor (702 vs 699 GB/s at (6912,1536); 369 vs 370 GB/s at (11008,4096)) and wide matrices cap at ~370-400 GB/s even for pure reads, so the reported M=8 -23% regression does not survive the cold regime
- update docs/developer/cuda_kernels.md accordingly

Benchmark: L20 (sm_89), PyTorch 2.11.0+cu128, rotated weight copies >= 240 MB to defeat the 96 MB L2; end-to-end through the built module at (6912,1536) reaches 738-752 GB/s for M in [1, 4] and 702 GB/s at M=8, about +8% at M=2/4 and +6% at M=8 over the removed warp path

docs/developer/cuda_kernels.md

@@ -125,12 +125,14 @@ BF16 inputs with `M` in `[1, 8]` and K divisible by 8. It preserves the BF16
 rounding boundaries of the two projection outputs, SiLU output, and final
 product while accumulating dot products in FP32.
 
-The kernel contains two output-row tilings. A CTA-reuse path reads each up/gate
-weight chunk once and applies it to all M rows. The native AstrAI 1B shape
-`(N,K)=(6912,1536)` uses one warp per decode row for M=2/4/8; on L20 this
-removes the shared reductions and barrier and reduces M=4 CUDA-Graph latency
-from 0.0324 ms to 0.0181 ms. Wider LLaMA/GPT-NeoX matrices keep CTA reuse,
-because duplicating their weight reads across row warps regressed 1.3-4.2%.
+The kernel is a single CTA-reuse tiling: one CTA per output column reads each
+up/gate weight chunk once and applies it to all M rows. Block size is 256
+threads for M in [1, 7] and 128 for M=8, where the shorter shared-memory
+reduction tree wins under cold-HBM decode traffic. An earlier per-shape
+`(6912,1536)` warp-per-row variant and its dispatch table were removed: HBM
+measurements with rotated weights showed the table was tuned against L2-cache
+regime timing and was up to 6% slower than CTA reuse at M=2/4; the kernel is
+bandwidth-bound, so finer variant selection is noise.
 
 Dense `MLP` modules route through the SwiGLU backend. `ASTRAI_SWIGLU=0` keeps
 the unfused linear backend, and `1` explicitly forces the fused primitive.

docs/developer/swiglu_benchmark.md

@@ -56,3 +56,28 @@ the existing unfused linear backend, including any independently qualified
 GEMV dispatch. `ASTRAI_SWIGLU=1` remains an explicit benchmark/experimentation
 switch for callers that accept normal BF16 reduction-order variation. A future
 automatic band must repeat both the performance and checkpoint-output gates.
+
+## HBM re-measurement and kernel simplification
+
+The operator numbers above are L2-resident: the AstrAI pair is 40.5 MB,
+smaller than the 96 MB L2, so a tight timing loop re-reads warm weights
+(13.75 us implies ~3.1 TB/s, far above the 864 GB/s spec). Real decode rotates
+~1 GB of per-layer weights through L2 every step, so every call is cold.
+
+Re-measuring with rotated weight copies (>= 240 MB working set) on the same
+L20 showed:
+
+- The fused CTA-reuse kernel sits at the dual-stream cold-read floor
+  (702 vs 699 GB/s at (6912,1536); 369 vs 370 GB/s at (11008,4096)). Wide
+  LLaMA matrices cap at ~370-400 GB/s regardless of kernel, even for a
+  pure-read loop, so the old per-variant gaps there were noise.
+- The `(6912,1536)` warp-per-row variant (formerly M=2/4/8) is 2-6% slower
+  than CTA reuse at M=2/4 under cold weights and no longer wins at M=8 once
+  the CTA drops to 128 threads. It and its dispatch table were deleted.
+- New rule: 256 threads for M in [1, 7], 128 threads for M=8. End-to-end
+  through the built module at (6912,1536): 738-752 GB/s for M in [1, 4] and
+  702 GB/s at M=8 (+6% over the removed warp path).
+
+The M=8 CUDA-Graph regression reported above (`-23.24%`) does not survive the
+cold-weight regime: cuBLAS reaches L2 bandwidth in the warm loop while both
+fused paths converge to the same HBM floor.